跳至主要内容

How We Keep Your Data Safe

A plain-words guide for anyone (technical or not) who is thinking about connecting their database to Dataira.

The big picture​

You connect your database. Your customers ask questions in plain language. Dataira answers with numbers and charts. For that to work, three things must be true:

  1. We can never change your data.
  2. One customer can never see another customer's data.
  3. Your password never sits around where someone could steal it.

Here is exactly how we make each one true. No hand-waving.

1. We can read. We cannot write. Ever.​

When you connect a database, we ask you for a special database account that has read-only permission. Think of it as a library card: you can look at books, you cannot edit them, and you certainly cannot burn the library down.

We do not just trust that you did this. When you connect, we check it ourselves:

And that is only the first lock. Every question a customer asks becomes a database query. Before any query runs, it must pass three separate checks:

  • It must be a pure read question (SELECT). Anything that writes, deletes, or changes data is thrown away before it ever reaches you.
  • It must have a safety limit, so no query can ask for unlimited rows.
  • It must be wrapped in a read-only transaction, so the database itself refuses to run it if it tries to write, like a car that physically cannot go into reverse.

Three locks, one door.

2. Your customers only ever see their own data​

Most SaaS companies keep all their customers in one database, with a column like customer_id saying which row belongs to whom.

When one of your customers opens the Dataira chat, your server tells our server: "this is customer 42." From that moment:

  • Every single query gets customer_id = 42 attached by our server, not by the AI, not by the browser. Even if the AI tried to "helpfully" query customer 43, the query is rejected before it runs.
  • This rule is checked by a machine, not by a promise. We parse every query and refuse any that dodges the filter. No clever tricks, no exceptions.
  • If we cannot prove the filter is there, the question is not answered. Failing closed is the default.

We test this adversarially: automated tests constantly try to sneak one customer into another customer's answers, and they must fail every time.

3. Your database password is locked away​

  • Your password is encrypted the moment we receive it, with a key that is unique to you. In our database it looks like gibberish, and it stays gibberish.
  • The connection between us and your database is always encrypted (TLS), so nobody can eavesdrop on the wire.
  • Even our own staff cannot read your password back. It is used to connect, nothing else.
  • If your database lives inside a private network, you don't even give us the password. Instead, you run our tunnel tool inside your network, and it holds the password. It never leaves your building. We can only ever ask it to run read-only questions.

4. What the AI says is never trusted blindly​

The AI reads your table names and schema descriptions to understand your business. Treat everything it "sees" as untrusted input, like a suspicious letter: it may read the letter, it may never follow instructions found inside it. Generated SQL always passes the machine checks from section 1 and 2 before running, no matter who or what wrote it.

5. You can leave, and you can watch everything​

  • Every question, every query, and every answer is written down in an audit log you can read. Nothing happens in the dark.
  • Every customer you register can be removed with one call, and their data access ends immediately, even for sessions that are already open.
  • Disconnect a database any time. Nothing about your data is needed for Dataira to keep working for everyone else.

Honest limits (read this part too)​

We believe you should know exactly where the boundaries are:

  • Database-native rules (like Postgres RLS) are yours. We enforce isolation in our layer and never weaken yours. If you have row-level security policies, they keep working, we just add ours on top.
  • Don't paste secrets into chats. Dataira is for asking questions about your data. Passwords, API keys, and private text do not belong in prompts.

Want to verify all this yourself?​

  • Run the two-customer isolation test on your own integration. It takes minutes.
  • Review the detailed technical controls in Security & Compliance.
  • Watch the audit log while a test customer asks a question. See exactly what ran, and what it touched.

Questions? Talk to us. Security questions from partners are answered first, always.